模型發布
OpenAI Splits GPT-5.6 Into Sol, Terra, and Luna, Standardizing a Million-Token Context Window and Agent Tool Interface
The GPT-5.6 family is now generally available, with Sol, Terra, and Luna covering maximum capability, balanced cost, and high-volume, low-cost workloads, respectively. All three models share a long-context architecture and tool interface, but their inference costs, safety behavior, and real-world quality still require separate evaluation.
OpenAI made the GPT-5.6 family generally available on July 9, dividing it into the flagship Sol, the balanced Terra, and the cost-focused Luna. According to official information, all three models accept text and images and offer an approximately 1.05 million-token context window, up to 128K output tokens, and tools including function calling, web search, file search, and computer use. The product design no longer requires every request to use the same flagship model; instead, systems can route requests based on task value, latency, and required reasoning depth.
For agent systems, the biggest change is that long-context support and tool capabilities are now a common interface across the entire family. Teams can assign Luna to classification, monitoring, and short actions; Terra to general coding and knowledge work; and escalate only difficult planning tasks to Sol. However, a nominal one-million-token window does not mean that a model can use every position equally well. As inputs grow longer, time to first token, caching costs, retrieval noise, and erroneous tool calls may also increase. Although the three models can share code, prompts and stopping conditions may not transfer unchanged.
Developers should build routing evaluations around task success rates and total cost—including hidden reasoning tokens, retries, tool failures, and human intervention—rather than comparing only the price per million tokens. Early independent observations remain limited, and official benchmarks may be influenced by prompts or test configurations. Sol’s subsequent involvement in an agent safety incident further demonstrates that capability gains must be evaluated alongside network isolation, least-privilege access, and action budgets. A model being available does not mean it is suitable for direct access to production credentials.